CrackNet: A transformer-based approach for detecting microcrack in photovoltaic panels based on electroluminescence images

Abstract

Microcracks in photovoltaic (PV) panels affect power generation efficiency and system safety. Traditional detection methods cannot accurately identify defects with complex shapes due to background noise and low computational efficiency. This paper presents a transformer-based semantic segmentation model, CrackNet, for detecting microcracks in electroluminescence (EL) images. We design a scale-aware dynamic dilated attention (SDDA) mechanism to improve the model’s ability to detect local details and determine global dependencies. We incorporate a dynamic upsampling operator (DySample) to replace dynamic convolutions with a point-sampling strategy, significantly reducing computational complexity and improving processing speed. We propose a preprocessing method for EL images to suppress grid line interference and improve detection accuracy. Experimental results on a custom PV microcrack dataset show that CrackNet achieves an average intersection over union (mIoU) of 80.45% and mean pixel accuracy (mPA) of 82.47%, significantly outperforming mainstream models, such as U-Net and DeepLab v3+. Moreover, CrackNet has higher parameter efficiency (41.72 M parameters) and computational performance (67.10 G FLOPs) than similar algorithms.